Dream to Remember: Continual Steering of Generalist Robot Policies with World Models

arXiv 2026 33.6 method

TLDR

D2R continually steers frozen generalist robot policies to new tasks while using a pretrained world model to rehearse prior tasks, preventing catastrophic forgetting.

Reasoning

The paper's main strength is its strong empirical validation on CALVIN and LIBERO using off-the-shelf policies and world models, directly addressing continual learning and forgetting. A limitation is that the method depends heavily on world-model faithfulness, and the truncated abstract leaves some implementation details and computational trade-offs unspecified.

Read-first score

Read-first score 33.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 25%

62.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 338.

Keyword Scores

world model

model-based reinforcement learning world model

world simulator

generative world model

world dynamics prediction

interactive world model

video world model

Deep Analysis

Innovations

Introduces Dream to Remember (D2R), which continually steers a frozen generalist robot policy toward new tasks while retaining earlier skills, avoiding catastrophic forgetting in both policy fine-tuning and shared steering modules.
Uses a frozen pretrained world model to rehearse prior tasks through imagined rollouts, anchoring the steering module to earlier behavior without a replay buffer or new data.
Supports two learning modes: learning new tasks entirely in imagination when the world model is faithful enough, otherwise learning in the environment while using the world model only to rehearse prior tasks.
Demonstrates continual steering with frozen off-the-shelf generalist policies and world models, including π0 and A2World, rather than requiring task-specific retraining or policy weight updates.

Methodology

D2R trains a steering module for a frozen generative generalist policy as tasks arrive sequentially. A frozen pretrained world model rehearses prior tasks through imagined rollouts to anchor the steering module to earlier behavior, without a replay buffer or new data. If the world model is accurate enough, new tasks are learned entirely in imagination; otherwise, tasks are learned in the environment while the world model is used only for rehearsal. The approach is evaluated on CALVIN, LIBERO-90 with π0 and A2World, and a real Franka robot.

Key Results

On CALVIN, D2R improves eight tasks sequentially entirely in imagination, raising average success from 63.7% to 74.8%, while LoRA fine-tuning on the same sequence drops to 11.7%. With frozen π0 and A2World, D2R learns unseen LIBERO-90 tasks in the environment while preserving π0's prior LIBERO performance, and on a real Franka robot it improves three tasks sequentially entirely in imagination while retaining earlier improvements.

Limitations

D2R can learn new tasks entirely in imagination only when the world model is faithful enough; otherwise it must learn in the environment.
The evaluated LIBERO-90 setup uses a world model that is not accurate enough for policy optimization, limiting fully imagined learning in that setting.

Resources

Paper